On the Marginal Value of Water for Hydroelectricity
|
|
- Joshua Price
- 6 years ago
- Views:
Transcription
1 Chapter 31 On the Marginal Value of Water for Hydroelectricity Andy Philpott Introduction This chapter discusses optimization models for computing prices in perfectly competitive wholesale electricity markets that are dominated by hydroelectric generation. We revisit the relationship between partial equilibrium in perfect competition and system optimization of a social planning problem, and we use the latter to show how perfectly competitive electricity prices correspond to marginal water values. Most industrialized regions of the world have over the last 20 years established wholesale electricity markets that take the form of an auction that matches supply and demand. The exact form of these auction mechanisms varies by jurisdiction, but they typically require offers of energy from suppliers at costs at which they are willing to supply, and clear a market by dispatching these offers in order of increasing cost. Dayahead markets, such as those implemented in most North American regions, seek to arrange supply well in advance of its demand so that thermal units can be prepared in time. Since the demand cannot be predicted with absolute certainty, these day-ahead markets must be augmented with balancing markets to deal with the variation in load and generator availability in real time. In this chapter, we study markets with hydroelectric generators that use reservoirs of stored water. Generators with hydroelectric reservoirs face an inventory problem. They would like to optimize the release of water from reservoirs to maximize profits using a stochastic process of prices, but this process is not known and must be deduced by each agent using current and future market conditions and hydrological models of future reservoir inflows. For an agent controlling releases from a hydroelectric reservoir, the marginal cost of supply in the current period involves some modeling of opportunity cost that includes possible high prices in future states of the world with low inflows. 21 The research described in this paper has benefited from discussions with Ziming Guan, Tony Downward, Geoffrey Pritchard, and Golbon Zakeri. 405
2 406 Chapter 31. On the Marginal Value of Water for Hydroelectricity In this chapter, we restrict our attention to a setting in which all agents are price-takers who do not act strategically. Under this assumption, it is well known that electricity prices for a single trading period and a single location can be computed as shadow prices from a deterministic convex economic dispatch (ED) model that maximizes total social welfare. In a stochastic setting, where there is a stochastic process of inflows that is known to all agents, and they maximize expected profit using the probability law determining these inflows, a competitive (partial) equilibrium will correspond to the water-release policy that maximizes the expected welfare of a social planner. A stochastic optimization (SO) for maximizing expected social welfare can therefore be used to estimate the marginal cost of electricity supply in future states of the world. These values are estimates of the wholesale electricity prices in these states. This approach is demonstrated in Section 31.2 using a simple model. In a system with a single storage reservoir, electricity prices can be shown to be equivalent to the expected marginal value of water in the reservoir, in other words the extra social benefit that would be obtained from the optimal deployment of an extra cubic meter of stored water. Observe that this value comes from the optimal marginal use of the water and so the planner must have an optimization problem in mind. Policy iteration algorithms based on an early paper by Stage and Larsson [1713] are often used in practice to compute these marginal values. In Section 31.3, we apply these methods to a simple example. Most hydroelectricity systems comprise several reservoir storages, and so the social planner s optimization problem becomes a stochastic dynamic programming (DP) problem with a matching number of state variables. When this number exceeds the acceptable state dimension for classical DP (about three or four), the optimization method used most in practice is stochastic dual dynamic programming (SDDP) due to Pereira and Pinto [1458]. In Section 31.4, we outline this method, and we apply it to an example problem from the New Zealand electricity system in Section The estimated prices from a case study in 2008 are then compared with their historical counterparts The Social Optimum We consider a social planning problem in an electricity system of hydroelectric generating plants fed by reservoirs (i ) that are subject to uncertain inflows over some decision horizon of T periods. The system also has generating plants ( j )thatrun on thermal fuel at some known generation cost, and consumer segments c.for simplicity we shall assume that all generating plants sell power into an unconstrained transmission grid so that consumer demand can be aggregated at a single location to which power is sent. In general, the reservoirs are connected in a set of river chains represented by a network of m nodes (reservoirs and junctions)and l arcs (canals or river reaches). The topology of the network can be represented by the m l incidence matrix A, where a ik = 1 if node i isthetailofarck, 1 if node i is the head of arc k, 0, otherwise. By adding dummy nodes if necessary, we can ensure that every pair of nodes is joined by at most one arc. If there are several river chains, then the network need not be connected.
3 Andy Philpott 407 We first describe a deterministic model over the planning horizon t = 1,2,...,T. We let u(t) be a vector of flow rates (cubic meters per period) in the arcs in the network in period t. Some arcs correspond to generating stations. Some might correspond to spill around a station. For each arc k = 1,2,...,l, we specify a generation function G k (u k ) that is zero when the arc does not contain a station, and otherwise is a strictly concave function of flow rate u k for arcs corresponding to stations, giving the energy output in a period when the flow rate is u k.letx (t) denote a vector of reservoir storages in each node at the beginning of period t and ω(t) be a vector of uncontrolled reservoir inflows (in cubic meters) that have occurred in period t. These define a water balance constraint x(t + 1)=x(t) Au(t)+ω(t), t = 1,2,...,T. The hydro generators are accompanied by thermal generators j that produce a vector v of energy that costs j C j (v j ) to produce, where C j is a strictly convex function. Demand of consumer segment c is represented by d c, where the strictly concave function D c (d c ) denotes the welfare accrued by consumer segment c. Note that the derivative D c represents the inverse demand curve for consumer c, whichis strictly decreasing by assumption. Demand must be satisfied in each time period, so, in a deterministic model, l G k (u k (t)) + v j (t) d c (t), t = 1,2,...,T. k=1 j c When inflows to the reservoirs are random variables, the sequence of time intervals 1,2,...,T is replaced by a stochastic process. We represent this by a scenario tree, as shown in Figure Here each scenario tree node n spans a period t (which in our models is a week) and corresponds to a realization ω(n) of reservoir inflows in that period. To account for different inflow sequences, all variables previously indexed by t are now indexed by n. Each node has probability φ(n), where the probabilities of nodes in a given period sum to one. We denote by n the unique parent node of node n and by n+ its set of children. The leaf nodes of the tree correspond to the end of the decision horizon (period T ). For each leaf node n, we specify a strictly convex future-value function V n (x(n)) that measures the total future welfare in state n of having reservoir levels x(n). We are now in a position to formulate a stochastic social planning (SSP) problem. This is SSP: min φ(n) C j (v j (n)) D c (d c (n)) n j c φ(n)v (x(n)) n subject to (s.t.) l G k (u k (n)) + v j (n) d c (n), k=1 j c n, x(n)=x(n ) Au(n)+ω(n), n, d(n) 0, v(n) 0, x(n), u(n), n.
4 408 Chapter 31. On the Marginal Value of Water for Hydroelectricity Figure Each node n spans a period and corresponds to a realization Ω(n) of reservoir inflows in that period. Here we assume that and are compact convex sets and SSP has a nonempty feasible region. We shall assume in what follows that all nonlinear constraints satisfy constraint qualifications to guarantee the existence of Lagrange multipliers. In particular, we have Lagrange multipliers φ(n) p(n) for each n so that we can solve SSP by minimizing the Lagrangian LSSP: min φ(n) p(n)d c (n) D c (d c (n) n c c + C j (v j (n)) p(n)v j (n) l V (x(n)) p(n)g k (u k (n)) j k=1 s.t. x(n)=x(n ) Au(n)+ω(n), n, d(n) 0, v(n) 0, x(n), u(n), n. LSSP separates by agent. The consumer solves CP(c): max φ(n)(d c (d c (n)) p(n)d c (n)) Each thermal generator solves TP(j): max n s.t. d c (n) 0. φ(n)(p(n)v j (n) C j (v j (n)) n s.t. v j (n) 0, and the hydro generators together solve l HP: max φ(n) p(n)g k (u k (n)) + φ(n)v (x(n)) n n k=1 s.t. x(n)=x(n ) Au(n)+ω(n), n, x(n), u(n), n.
5 Andy Philpott 409 If all the river chains are operated by a single agent, then HP represents the problem of maximizing this agent s expected revenue and residual value given prices p(n) in each state of the world. If there is more than one hydro agent and HP is not separable by agent, then we require an additional condition to establish a correspondence between competitive equilibrium and an optimal social plan. Let denote the set of hydro agents and x i the vector of water stocks for reservoirs operated by agent i.werequire the existence of functions V i (x i ) so that V (x(n)) = i V i (x i (n)). Even with this condition, to enable a separation into agent problems, in general we require a price φ(n)π(n) for the flow conservation constraint. This gives a Lagrangian for HP, which is (x, u,π)= l φ(n) p(n)g k (u k (n)) + φ(n) V i (x i (n)) n k=1 n i + φ(n)π(n) (x(n ) Au(n)+ω(n) x(n)), n that we maximize over x(n), u(n). (x, u,π) can be rearranged to give (x, u,π)= l φ(n) p(n)g k (u k (n)) π(n) Au(n) n k=1 + m n+ φ(n)x(n) φ(m)π(m) π(n) φ(n) n \ + φ(n) (V i (x i (n)) π i (n)x i (n)) n i + n φ(n)π(n) ω(n). The values of π(n) are water prices, often called marginal water values. Attheendof the planning horizon, if x i (n) is not at a bound, then we require for a maximum that V i (x i (n)) = π i (n), n. For flow u k (n) through station arc k from h to i,wehave π(n) Au(n)=(π i (n) π h (n)) u k (n), so the owner of station k attempts in state of the world n to maximize l p(n)g k (u k (n)) + (π i (n) π h (n)) u k (n). k=1 In other words, he/she maximize revenue from u k (n) at price p(n), while paying π h (n) for upstream water and getting paid π i (n) for the water supplied to the downstream reservoir. At the optimal solution u k,wehave π h (n) π i (n)= p(n)g k u k (n), (31.1)
6 410 Chapter 31. On the Marginal Value of Water for Hydroelectricity which relates the spot price of energy p(n) to the marginal water value difference and the marginal conversion factor from water to energy. A reservoir owner in state of the world n makes x(n) as large as possible when the expected marginal water value at the next time period exceeds that at the current period and makes x(n) as small as possible when the expected marginal water value at the next time period is lower than that at the current period. These prices in equilibrium enable different owners on the river to extract the maximum welfare from the water. If there is no possibility of trading in water at market prices, then the market is incomplete, and the competitive equilibrium may not be the same as the optimal social plan. This observation is well known and is studied by Lino et al. [1198]. Henceforth we will assume that all the reservoirs and generating stations on a given river chain are owned and operated by the same generator, and that all the river chains have separate catchments. This means that HP separates into independent river chains each operated by a single agent. Hydro agent i solves HP(i): max l (i) φ(n) p(n)g k (u k (n)) + φ(n)v i (x(n)) n n k=1 s.t. x(n)=x(n ) Au(n)+ω(n), n, x(n), u(n) 0, where now x(n)(ω(n)) denotes the vector of reservoir storages (inflows) on river chain i and u(n) l (i) the vector of flows through the l (i) arcs in the river chain. This defines a perfectly competitive equilibrium defined by the individual optimality conditions and market clearing condition: CE: u i (n), x i (n) argmaxhp(i), v j (n) argmaxtp(j), d c (n) argmaxcp(c), 0 i l (i) G k u i k (n) + v j (n) d c (n) p(n) 0. j c k=1 By comparing optimality conditions, we have the following result. Proposition Suppose all the reservoirs and generating stations on any given river chain are owned and operated by the same generator and V (x)= i V i (x i ). The (unique) welfare-maximizing solution to SSP is the same as the (unique) competitive equilibrium solution to CE. Each model HP(i) can be solved by computing an optimal water release policy to maximize the expected revenue at energy prices p(n). The optimal policy will generate Lagrange multipliers π(n) for the water conservation constraints. These will be related to the energy prices through (31.1). When each hydro agent controls a single
7 Andy Philpott 411 reservoir, then it is tempting to suppose that its optimal water release policy can be computed by price decomposition and DP. Unfortunately, this is not always straightforward because the energy prices p(n) are not guaranteed to be stagewise independent even if the inflows are independent [166]. On the other hand, DP can be applied to the system optimization problem as long as the number of reservoirs (corresponding to states) is not too large. Given a system optimal solution to the social planning problem, it is possible to derive the competitive equilibrium prices and agent policies using the above proposition. In the next section, we study the computation of competitive equilibrium prices for the simplest possible system, with one hydro agent and one reservoir. We then show how models with several reservoirs might be solved using SDDP [1458, 1671]. We conclude with some results of estimating competitive prices in the New Zealand wholesale electricity market A Single-Reservoir Model One of the first models for reservoir water valuation was developed by Stage and Larsson [1713]. It seeks to compute marginal water values w i (t) corresponding to discrete reservoir levels i = 0,1,...,N using a form of policy iteration. When the reservoir is full (i = N), the marginal value of water is set to zero (here we assume that surplus water can be spilled with no penalty). When the reservoir is empty (i = 0), the marginal value of water is set to some high value that we denote as C. Typically this will be a shortage cost, also known as the value of lost load. Note that this convention implies that the marginal value at level i is the directional derivative w i = W i W i+1, where W i is the average expected cost of meeting demand in each future period if the current reservoir level is i.ifw i is linearly interpolated, then we obtain a nonsmooth convex function. This implies W i 1 W i = w i 1 w i = W i W i+1, with strict inequality occurring as the rule. The expected marginal value of water at level i = j,say,isdeterminedin[1713] by constructing sample paths of reservoir levels that would be produced by an optimal release policy (say to meet a known demand) under random sequences of inflows. Starting at level i = j at time of year t, s out of S such sample paths will reach i = 0 before i = N. The expected marginal water value w j (t) is then estimated to be s S C. This process is repeated for all starting levels and times. Although the Stage and Larsson method is quite general, we will examine it here as applied to a specific problem instance. We assume that reservoir inflows are stagewise independent and have the same distribution at each time of the year. Thus we can dispense with the time index in the formulation and seek values w i. To simplify the analysis further we will assume that the reservoir has inflow in each period that equals twowith probability p and zero with probability (1 p). Demand for energy requires that we release exactly one unit of water in each period. Thus in each time step the reservoir level is a simple random walk, increasing by one unit with probability p and decreasing by one unit with probability 1 p. Observe that there is no policy choice since the water usage is prescribed by demand. In this setting, the Stage and Larsson method amounts to iterating toward the solution of a fixed-point problem.
8 412 Chapter 31. On the Marginal Value of Water for Hydroelectricity In other words, we obtain the following procedure: Set ν = 0, repeat w ν i = 0; for i = 0toN do if i = 0, set w ν+1 = C ; i if i = N, set w ν+1 = 0; i if 0 < i < N, set w ν+1 = (1 p)w ν i i 1 + pwν i+1 ; end for; set ν := ν + 1 until w ν+1 = w ν. It is easy to see that this procedure converges to the fixed point defined by w i =(1 p)w i 1 + pw i+1, i = 1,2,...,N 1, w 0 = C, w N = 0. (31.2) In fact this simple model can be solved exactly for w because there is a closed-form solution for the probability that any random walk starting at i = j will hit barrier i = 0 before it hits barrier i = N (see, e.g., [854]). When p = 0.5, this gives w j = C N j Otherwise, letting r = 1 p p,weget 1 r C w j = 1 r r j + r j r N 1, j < N, N 0, j = N. The solution corresponds to the fixed-point calculation N. η + W 0 = C +(1 p)(w 0 )+ pw 2, η + W i =(1 p)w i 1 + pw i+1, W N+1 = W N, i = 1,2,...,N, (31.3) for an average cost per period Markov decision process, where η is the long-run average cost per period of meeting demand. Given a solution W to (31.3), it is easy to see that setting w i = W i W i+1 solves (31.2). An alternative model uses discounting. Suppose costs are discounted in each period with rate α, giving a discount factor γ = 1. This gives a slightly different fixed-point 1+α problem: w i =(1 p)γ w i 1 + pγ w i+1, i = 1,2,...,N 1, w 0 = C, w N = 0. (31.4) The discounted solution corresponds to the fixed-point calculation W 0 = C + γ ((1 p)w 0 + pw 2 ), W i = γ (1 p)w i 1 + pw i+1, W N+1 = W N. i = 1,2,...,N, (31.5) Here W j is the expected discounted cost of meeting demand starting from reservoir level i = j. Given a solution W to (31.5), setting w i = W i W i+1 solves (31.4). In both of these models there is an implicit assumption on how shortage costs are incurred. When the reservoir level hits zero, it is assumed that the cost C is incurred immediately. This assumes that none of the inflow occurring over the next period
9 Andy Philpott 413 can be used to cover demand. In other words, this is a here-and-now or decision-hazard problem, which is modelled using (31.3) or (31.5). In contrast, one might use a wait-and-see or hazard-decision assumption. Given a reservoir level i = j, the random inflow is observed and used with reservoir water to meet demand if possible. Inflow that cannot be stored is spilled. Thus, when the reservoir is empty, with probability 1 p we must meet demand with no inflow, incurring acostofc, resulting in an empty reservoir, or we meet demand with an inflow of two units, leaving the reservoir with one unit in it. When the reservoir is full and inflow is two, we can use half of it and leave the reservoir full. If the inflow is zero, then the reservoir moves to state N 1. The fixed-point problem (31.3) is now η + S 0 =(1 p)(c + S 0 )+ ps 2, η + S i =(1 p)s i 1 + ps i+1, S N+1 = S N, i = 1,2,...,N, (31.6) and (31.5) becomes S 0 =(1 p)(c + γ S 0 )+ pγ S 2, S i =(1 p)γ S i 1 + pγ S i+1, S N+1 = S N, i = 1,2,...,N, (31.7) where in (31.6) S j is the expected (wait-and-see) discounted cost of meeting demand starting from reservoir level i = j, and in (31.7) S j is the corresponding value in the average cost per period model Thermal Generation The Stage and Larsson methodology can be extended to a setting where the policy being simulated generates energy using thermal plant (at some cost) as well as using releases from the hydro reservoir. This is the basis for a number of reservoir optimization models in practical use (see, e.g., [326]). It is important to observe that the values w i derived from such a simulation will only be an expected marginal value for an optimization problem, if the policy being simulated is indeed optimal for that problem. In other words, the expected marginal value of water is the expected extra benefit that accrues from using this water in the most beneficial way. The optimal policy ultimately requires the solution of a stochastic dynamic optimization problem. To include thermal generation in our example, we alter demand so that it is two units per period and suppose there is a thermal generator that costs c < C to produce one unit of energy. Thus, to meet demand, we can either release two units of water or release one unit of water and run the thermal plant. The thermal generator is assumed to be flexible in that it can be switched on and off with no penalty. It is typical to specify an optimal policy in terms of a marginal water value as follows. If c < w i,then we run the thermal plant. If c > w i, then we release water. If c = w i, then the policy can be any combination of thermal generation and water release. In this model, the Stage and Larsson methodology (with discounting) might yield the following marginal water value algorithm:
10 414 Chapter 31. On the Marginal Value of Water for Hydroelectricity Set ν = 0, repeat w ν i = 0; for i = 0toN do if i = 0, set w ν+1 = C ; i if i = N, set w ν+1 = 0; i if 0 < i < N, if w ν < c, wν+1 =(1 p)γ w ν i i i 2 + pγ wν i ; if w ν i > c, wν+1 =(1 p)γ w ν i i 1 + pγ wν i+1 ; end for; set ν := ν + 1; until w ν+1 = w ν. If we relax the integrality constraint so that the thermal plant can generate any amount between zero and one, then the marginal water value will equal c when this plant is partially dispatched. There is a range of fractional water levels between x and x + 1, say, for which this will be the optimal action. For almost every instance of this relaxed problem, the marginal value of water will equal c at exactly one integer level i and be different from c for other levels. The correct stationarity conditions (for any choice of γ) that correspond to the minimum cost solution are w 0 = C, w i = γ (1 p)w i 1 + pw i+1, i = 1,...,k 1, w k = c [w k+1, w k 1 ], i = k, w i = γ (1 p)w i 2 + pw i, i = k + 1,...,N 1, w N = 0. (31.8) Thecorrectvalueofk can be found by computing an optimal policy. Let W i denote the expected future discounted cost of meeting demand. The here-and-now fixed-point problem (with discounting) is W 0 = C + c + γ ((1 p)w 0 + pw 2 ), W 1 = c + γ ((1 p)w 0 + pw 2 ), W i = min{c + γ (1 p)w i 1 + pw i+1,γ (1 p)wi 2 + pw i }, i = 2,...,N, W N+1 = W N. (31.9) The optimal solution gives a threshold type of policy, i.e., we release two units of water if γ (1 p)w i 2 + pw i < c + γ (1 p)wi 1 + pw i+1 or c >γ (1 p)w i 2 + pw i. The right-hand side is the expected discounted marginal water value if we meet demand from hydro only, and water is released if this is less than the cost c of thermal generation. An optimal policy can be found using linear optimization (LO) and determines a threshold index k below which the thermal plant is always run at capacity and above which it is not. Given this value of k, the optimal policy satisfies W 0 = C + c + γ ((1 p)w 0 + pw 2 ), W 1 = c + γ ((1 p)w 0 + pw 2 ), W i = c + γ (1 p)w i 1 + pw i+1, W i = γ (1 p)w i 2 + pw i, W N+1 = W N. i = 2,...,k, i = k + 1,...,N
11 Andy Philpott 415 By setting w i = W i W i+1, we can derive the marginal values w that will hold at optimality. These satisfy w 0 = C, w i = γ (1 p)w i 1 + pw i+1, i = 1,...,k 1, w k = c [w k+1, w k 1 ], w i = γ (1 p)w i 2 + pw i, i = k + 1,...,N 1, w N = 0, (31.10) which are the same equations as (31.8). The case without discounting is similar. If we form an average cost per period Markov decision problem, then we seek the solution to max g W 0 = η + C + c +(1 p)w 0 + pw 2, W 1 = η + c +(1 p)w 0 + pw 2, W i η + c +(1 p)w i 1 + pw i+1, W i η +(1 p)w i 2 + pw i, W N+1 = W N. i = 2,...,N, i = 2,...,N, (31.11) As before, the optimal policy can be found using LO and determines a threshold index k below which the thermal plant is always run at capacity and above which it is not. Given this value of k, we can derive the marginal values w that will hold at optimality. This gives the system w 0 = C, w i =(1 p)w i 1 + pw i+1, i = 1,...,k 1, w k = c, w i =(1 p)w i 2 + pw i, i = k + 1,...,N 1, w N = 0, (31.12) which is the same as (31.8) when γ = 1. In fact in the average cost per period Markov decision problem for this system, k turns out to be always equal to N 1, and C, i = 0, ci+c (N 1 i), i = 1,...,N 2, w i = N 1 (31.13) c, i = N 1, 0, i = N, when p = 0.5 and w i = C, i = 0, C 1 r 1 r r i + r i r N 2 N 1 1 r, i = 1,...,N 2, +c 1 r 1 + r + + r i 1 N 1 c, i = N 1, 0, i = N, (31.14) otherwise. Proposition IfC > c, then there is a unique monotonic solution to (31.12) defined by choosing k = N 1.
12 416 Chapter 31. On the Marginal Value of Water for Hydroelectricity Proof. If k < N 1, then yielding w k+1 = (1 p)w k 1 + pw k+1, w k+1 = w k 1, and hence by monotonicity of w i, we obtain Solving for i = 1,...,k 1gives w k 1 = w k+1 = w k = c. w i =(1 p)w i 1 + pw i+1 w i = c, i = 0,...,k + 1, contradicting w 0 = C > c.thusk = N 1, and the unique solution is given by (31.13) if p = 0.5, or (31.14) if p Multiple-Reservoir Models In most real applications, marginal water values are obtained from the optimal operation of many reservoirs, sometimes linked in a cascaded river system, and almost always linked by some electricity transmission system. When the number of reservoirs is larger than two or three, representing their optimal operation using Markov decision problems becomes more difficult, due to the curse of dimensionality. Approximations of various forms are necessary. The approximation in widespread use is based on the SDDP algorithm of Pereira and Pinto [1458]. Here the marginal values of water emerge from an approximate stage problem that is modeled as an LO problem. This method can be shown to converge to an optimal policy almost surely under mild assumptions on the sampling process (see [804, 1471]). The subgradients of the optimal value function define the marginal water value. The class of problems we consider have T stages, denoted by t = 1,2,...,T, in each of which a random right-hand side vector b t (ω t ) m has a finite number of realizations defined by ω t Ω t. We assume that the outcomes ω t are stagewise independent, and that Ω 1 is a singleton, so the first-stage problem is z = min c 1 x 1 + [Q 2 (x 1,ω 2 )] s.t. A 1 x 1 = b 1, x 1 0, (31.15) where x 1 n is the first-stage decision, and c 1 n is a cost vector, A 1 is an m n matrix, and b 1 m. We denote by Q 2 (x 1,ω 2 ) the second-stage costs associated with decision x 1 and realization ω 2 Ω 2. The problem to be solved in the second and later stages t, given decisions x t 1 and realization ω t, can be written as Q t (x t 1,ω t )=min ct x t + [Q t+1 (x t,ω t+1 )] s.t. A t x t = b t (ω t ) E t x t 1 [π t (ω t )], x t 0, (31.16)
13 Andy Philpott 417 where x t n is the decision in stage t, c t its cost, and A t and E t m n matrices. Here π t (ω t ) denotes the dual variables of the constraints. In the last stage we assume either that [Q T +1 (x T,ω T +1 )] = 0 or that there is a convex polyhedral function that defines the expected future cost after stage T. For all instances of (31.16), we assume relatively complete recourse, whereby (31.16) at stage t has a feasible solution for all values of x t 1 that are feasible for the instance of (31.16) at stage t 1. Relatively complete recourse can be ensured by introducing artificial variables with penalty terms in the objective. The SDDP algorithm performs a sequence of major iterations, each consisting of a forward pass and a backward pass through all the stages, to build an approximately optimal policy. In each forward pass, a set of N scenarios is sampled from the scenario tree and decisions are made for each stage of those N scenarios, starting in the first stage and moving forward up to the last stage. In each stage, the observed values x t (s) of the decision variables x t, and the costs of each stage in all scenarios s,aresaved. The SDDP algorithm builds a policy that is defined at stage t by a polyhedral outer approximation (OA) of [Q t+1 (x t,ω t+1 )]. Thisapproximationisconstructed using cutting planes called Benders cuts, or just cuts. In other words, in each tth-stage problem, [Q t+1 (x t,ω t+1 )] is replaced by the variable θ t+1, which is constrained by a set of linear inequalities θ t+1 ḡ t+1,k,s x t h t+1,k,s, k = 1,2,...,K, s = 1,2,...,N, (31.17) where K is the number of backward passes that have been completed and ḡ and h are defined by (31.20) and (31.21) below. With this approximation, the first-stage problem is z = min c 1 x 1 + θ 2 s.t. A 1 x 1 = b 1, θ 2 ḡ 2,k,s x 1 h 2,k,s, x 1 0, and the tth stage problem becomes k = 1,2,...,K, s = 1,2,...,N, (31.18) Q t (x t 1,ω t )=min ct x t + θ t+1 s.t. A t x t = b t (ω t ) E t x t 1, [π t (ω t )] θ t+1 ḡ x t+1,k,s t h t+1,k,s, k = 1,2,...,K, s = 1,2,...,N, x t 0, (31.19) where we interpret the set of cuts as being empty when K = 0. At the end of the forward pass, a convergence criterion is tested, and if it is satisfied then the algorithm is stopped; otherwise it starts the backward pass, which is defined below. In the standard version of SDDP (see [1458]), the convergence test is satisfied when z, the lower bound on the expected cost at the first stage, is statistically close to an estimate of the expected total operation cost obtained by averaging the cost of the policy defined by the cuts when applied to the N sampled scenarios. In this simulation, the total operation cost for each scenario is the sum of the present cost (c t x t )overall stages t and any end-of-horizon future cost. If the convergence criterion is not satisfied, then SDDP amends the current policy using a backward pass that adds N cuts to each stage problem, starting at the
14 418 Chapter 31. On the Marginal Value of Water for Hydroelectricity penultimate stage and working backward to the first. To compute the coefficients for the cuts, we solve the next-stage problems for all possible realizations (Ω t+1 ) in each stage t and scenario s. The cut for (31.19), the tth (approximate) stage problem in scenario s, is computed after its solution x t k (s) has been obtained in the forward pass immediately preceding the backward pass k. Solving the (t + 1)th (approximate) stage problem for every ω t+1 Ω t+1 gives π t+1,k,s = [π t+1 (ω t+1 )], which defines the cut gradient ḡ t+1,k,s = π t+1,k,s E t+1 (31.20) and its intercept ht+1,k,s = [Q t+1 ( x k t (s),ω t+1 )] + π t+1,k,s E t+1 x k t (s). (31.21) The SDDP algorithm is initialized by setting θ t =, t = 2,...,T, K = 0, k = 1. Thereafter, the algorithm performs the following three steps repeatedly until the convergence criterion is satisfied: 1. Forward Pass For t = 1, solve (31.18) and save x k 1 (s)=x 1, s = 1,...,N,and z k = z; For t = 2,...,T and s = 1,...,N, Solve (31.19) setting x t 1 = x k t 1 (s), and save x t k (s) and Q t ( x k t 1 (s),ω t ). 2. Standard Convergence Test (at 100(1 α)% confidence level). N T Calculate the upper bound: z u = 1 c N t x t k(s), s=1 t=1 1 N T 2 σ u = c N t x t k(s) zu. 2 s=1 t=1 Calculate the lower bound: z l = z k ; Stop if z l > z u Z α 2 N σ u, where Z α is the (1 α) quantile of the standard normal distribution; otherwise go to the backward pass. 3. Backward Pass For t = T,...,2,and s = 1,...,N, For ω t Ω t, solve (31.19) using x k t 1 (s) and save π t,k,s = [π t (ω t )] and Q t ( x k t 1 (s),ω t ); Calculate the kth cut for s in stage t 1 using (31.20) and (31.21). Set K = K + 1, k = k + 1. SDDP terminates when the estimated value z u of the current policy is closer than Z α 2 σ N u to a lower bound z l on the optimal value. The expected value of an optimal
15 Andy Philpott 419 policy for the model (with finitely many inflow outcomes) can then be assumed to lie in the interval [z l, z u + Z α 2 σ N u ] with probability at least (1 α). Energy prices are deduced from the solution to SDDP by simulating a single stage of the (approximately) optimal policy (using cutting planes to represent the Bellman function) and recording the shadow prices on the demand constraint. Marginal water values in each reservoir at level x are determined from the subdifferential of [ Q t (x t 1,ω t )] evaluated at x. Often this will be a singleton defined by the coefficients of the highest cut at x. It is important to observe that the above algorithm operates under a wait-and-see assumption. In other words, we solve (31.19) with full knowledge of the current period s inflow, which can be used to generate electricity to satisfy demand. This means that shortage costs might be avoided even when all reservoir levels are at their minimum (see Section 31.2 above). Since shortages arising in simulations of an optimal policy yield marginal water value estimates, the wait-and-see assumption is likely to underestimate prices. Based on the single-reservoir analysis, an optimal generation schedule will use a merit order of water release and thermal generation based on marginal water value and marginal cost. With linear dynamics and with linear production functions, this produces a bang-bang type of control policy. It is well known, however, that many such systems admit singular control (i.e., one not uniquely determined by the adjoint solution). For example, the separated continuous LO problems discussed in [81] have this property. Examples in a hydro-scheduling context are given in [1468] Application We now show how SDDP can be applied to estimate wholesale prices for the New Zealand electricity system under the assumption of perfectly competitive risk-neutral agents. We begin by solving a model in which the inflows are assumed to be stagewise independent. We use the DOASA software described in [1473]. We seek to estimate New Zealand wholesale electricity prices for the week beginning March 10, The physical characteristics of the system are described in [1473] and [1472] (which together provide a more detailed study of the New Zealand wholesale electricity market over the period ). The electricity network shown in Figure 31.2 is approximated by a four-reservoir system by aggregating the storage reservoirs in the Waitaki river system into a single lake (Lake Waitaki). The resulting Waitaki system is shown in Figure Lake Ohau (OHU), which has limited storage capacity, is now treated as a runof-river resource, and reservoirs Benmore (BEN), Pukaki (PKI), and Tekapo (TEK) are treated as a single lake. Since these occur in a cascade in the river system, the aggregation of water in each must account for their specific energies. Thus Benmore and Pukaki storage and inflows are discounted by their specific energies before being added to the aggregate Lake Waitaki, which has the same specific energy as Tekapo. To allow for the storage ability of Benmore and Pukaki, Lake Waitaki water can be restored to these lower lakes through the arcs connecting Lake Waitaki to its children. These arcs have multipliers that inflate the water volume transferred to correspond to the difference in specific energies of Lake Waitaki and the lower lakes. The transmission network is approximated by three nodes, one for the South Island and two for the North Island. Limits on power transfer between nodes correspond to reported transmission limits. Energy deficit in any stage is met by load shedding at an
16 420 Chapter 31. On the Marginal Value of Water for Hydroelectricity Figure Approximate network representation of New Zealand electricity network showing main hydroelectricity generators. Reprinted with permission from Elsevier [1473]. increasing shortage cost in three tranches. This is equivalent to having three dummy thermal plants at each node with capacities equal to 5% of load, 5% of load, and 90% of load, for each load sector, and costs as shown in Table Load shed above 10% of demand costs a value of lost load of $20,000/MWh, which is roughly equivalent to five hours of load curtailment per year assuming annual amortized peaking-plant costs of $100/kW. Table Load reduction costs (NZD/MWh) and proportions of industrial load, commercial load, and residential load in each island. Up to 5% 5% to 10% Above 10% North Is. South Is. Industrial $1,000 $2,000 $20, Commercial $2,000 $4,000 $20, Residential $2,000 $4,000 $20, We choose historical inflows from each week in the years to give an empirical inflow distribution with 10 outcomes per stage. We assume that these are stagewise independent. The first model attempts to compute an optimal policy for 52 weeks starting at midnight on March 10, The computed electricity prices for the week beginning March 11, 2008, are the same at each node. They are $38.45/MWh, $38.76/MWh, and $39.84/MWh, corresponding to off-peak, shoulder, and peak periods. We can simulate the corresponding policy using inflows sampled from the assumed stagewiseindependent distribution. This gives storage levels in Lake Waitaki from 100 random sequences, as shown in Figure If we simulate the same policy with the 30 most recent historical inflow sequences, then we obtain storage levels in Lake Waitaki as shown in Figure The storage levels shown hit both extremes more often than the simulation that assumes stagewise independence. A heuristic that has been suggested to overcome the optimistic biases from assuming stagewise independence is called inflow spreading [1548]. Assuming independence over k weeks results in a lower variance in total inflow over this period than the data
17 Andy Philpott 421 Figure Aggregated model of Waitaki system with storage lakes Pukaki, Tekapo, and Benmore combined to make Lake Waitaki. Their inflows (scaled by specific energies) combine to produce an inflow to Lake Waitaki. support. Inflow spreading modifies the inflows using a heuristic that increases the variance of assumed independent inflows so that the variance of k consecutive independent inflows matches the observed variance of this sum. Consider historical inflow sequence h(t, y) of inflows in week t of year y = 1,2,..., N, and suppose k = 4. Let and w(t, y)=h(t, y)+h(t + 1, y)+h(t + 2, y)+h(t + 3, y), a(t)= y h(t, y), b(t)= N d(t, y)=max 0,a(t)+ The model then uses adjusted inflows ω(t, y)= y w(t, y), N (w(t, y) b(t)) 2. Nd(t, y)a(t), t = 1,2,...,52, y = 1,2,...,N. y d(t, y)
18 422 Chapter 31. On the Marginal Value of Water for Hydroelectricity Figure Simulated volumes in Lake Waitaki reservoir for 100 randomly generated inflow sequences. Figure Simulated volumes in Lake Waitaki reservoir for historical inflow sequences An optimal policy for 52 weeks starting at midnight on March 10, 2008, can be computed using adjusted inflows with k = 4. We can simulate the corresponding policy using inflows sampled from the assumed stagewise-independent distribution. This gives storage levels in Lake Waitaki from 100 random sequences (with spread inflows)
19 Andy Philpott 423 Figure Inflow spreading policy simulated volumes in Lake Waitaki reservoir for 100 randomly generated inflow sequences. Figure Inflow spreading policy simulated volumes in Lake Waitaki reservoir for historical inflow sequences as shown in Figure If we simulate the same policy with the 30 most recent historical inflow sequences, then we obtain storage levels in Lake Waitaki as shown in Figure The stock levels hit zero in only one year out of 30. The electricity prices
20 424 Chapter 31. On the Marginal Value of Water for Hydroelectricity for the week beginning March 11, 2008, are computed and are the same at each node. They are $40.10/MWh, $40.81/MWh, and $41.02/MWh, corresponding to off-peak, shoulder, and peak periods. The inflow spreading has resulted in slightly higher prices. The prices estimated by our model have been computed using SDDP, which makes use of a sampling procedure. This means that there will be some sampling error in their estimation. We first attempt to estimate this error by simulating the policy that we have obtained with perturbed initial reservoir levels. A simulation of the optimal (inflow spreading) policy over 2,500 random inflow sequences yields a sample average optimal cost of $354,861,576. Increasing the initial reservoir volume by 1 million cubic meters and simulating with common random numbers yields a sample average optimal cost of $354,812,679. The difference in costs is $48,897, with a standard error of $2,685. Assuming that the SDDP policy is optimal yields an estimate of marginal water value for Lake Waitaki of $0.0489/m 3, with a 95% confidence interval [0.0435, ]. To convert a cubic meter to an energy value, we must divide by the specific energy of Lake Waitaki, which is MWh/m 3, giving a 95% confidence interval for the perfectly competitive wholesale Benmore price of [$38.46, $48.01] Discussion One can compare the computed estimates of weekly prices with observed prices in the New Zealand wholesale market in the week beginning March 11, These range from $60/MWh to $190/MWh. The marginal water values computed by our SDDP algorithm are significantly lower than those reflected in observed electricity prices. There are several possible causes of this. 1. Poor representation of stagewise dependence in inflows: We have represented stagewise dependence in inflows using an inflow spreading heuristic based on the approach currently used by some New Zealand electricity generators. We have also ignored the melting of snowpack, which affects the inflow dynamics in the Waitaki catchment. The true reservoir inflow processes in New Zealand exhibit a more complicated long-range dependence than we have in our model. New inflow models that aim to improve the modeling of this dependence are discussed in [1518]. 2. Wait-and-see representation of uncertainty: Our SDDP model assumes that inflows are observed in any week before release decisions are made for that week. As discussed above, this assumption implies that the model will underestimate the probability of a shortage event, which will underestimate marginal water values. On the other hand, local meteorologists make reasonably reliable weather forecasts for the next three to five days, so we conjecture that the wait-and-see model should give reasonably accurate estimates, all else being equal. 3. Assumptions about constant fuel costs not reflecting fuel contract structure: In the New Zealand wholesale market, all large electricity generators (including thermal generators) offer increasing supply functions to the system operator. We have assumed that thermal short-run marginal cost is constant for each plant, so these supply functions in our model are assumed to be constant. Increasing supply functions for thermal plant are often interpreted as evidence of market power exercise, but they have alternative explanations. The owners of these plants typically operate under fuel contracts that affect the cost of their day-to-day fuel consumption. For example, coal is stockpiled, and so if replenishment is expensive
21 Andy Philpott 425 and time-consuming, its marginal value increases as a stock is depleted. On the other hand, gas is typically supplied under take-or-pay contracts over a contract period of several months. This effectively makes above-average consumption in any period more costly as it incurs a risk of exceeding the contract volume and having to pay a penalty. 4. Agents making different assumptions about the probabilities of dry inflow outcomes: We have assumed that all agents have the same probability distribution of inflow outcomes, sampled with equal probability from historical records. Recent climate observations (such as El Nino Southern Oscillation forecasts) might affect agents subjective assessments of these probabilities. 5. Agents acting strategically in offering generation at higher than competitive prices: We have assumed that all agents act as price-takers. This is not required by the rules governing the operation of the New Zealand wholesale electricity market. Some increase in price might be attributed to agents behaving strategically. 6. Agents not being risk neutral: We have assumed that all agents are risk neutral, and so they act to maximize expected profits. However, generators and purchasers are generally risk averse, so there is a possible discrepancy between a social optimum (and corresponding marginal water values) and a competitive partial equilibrium. This discrepancy is explored in detail in [1470],where competitive agents facing uncertain reservoir inflows are endowed with possibly different coherent time-consistent risk measures. In [1470], it is shown that if enough contracts are made available to agents to trade, and they are not too different in their attitudes to risk, then a social risk measure emerges in equilibrium that can be used to compute a risk-averse social plan that coincides with the partial competitive equilibrium. The risk-averse social plan can be computed using a risk-averse version of SDDP (see [1469]). In this setting, prices will increase as agents trade off their current costs against risk-adjusted future costs, which will make them more conservative than a risk-neutral agent when water is scarce. If the market for risk is incomplete, so thermal plant owners and hydro plant owners have no ability to hedge their different risk positions, then a perfectly competitive equilibrium need not correspond to a social planning solution, even if the planning solution is risk averse. Of course, even if we assume completeness in markets for risk, agents in hydro-dominated electricity markets face a plethora of commercial risks apart from uncertain inflows (which are the sole source of uncertainty in our model). It is conceivable that these other uncertainties are regarded as more critical than inflow risk, influencing the price markups we observe in the week beginning March 11, Identifying which combination of these factors is driving the difference between modeled and observed prices is difficult. However, for regulators of hydro-dominated electricity markets who monitor inefficiencies arising from imperfect competition or market incompleteness, this is a critically important task. SO models and stochastic equilibrium models are powerful tools to enable this to happen.
22
Worst-case-expectation approach to optimization under uncertainty
Worst-case-expectation approach to optimization under uncertainty Wajdi Tekaya Joint research with Alexander Shapiro, Murilo Pereira Soares and Joari Paulo da Costa : Cambridge Systems Associates; : Georgia
More informationInvestigation of the and minimum storage energy target levels approach. Final Report
Investigation of the AV@R and minimum storage energy target levels approach Final Report First activity of the technical cooperation between Georgia Institute of Technology and ONS - Operador Nacional
More informationOn solving multistage stochastic programs with coherent risk measures
On solving multistage stochastic programs with coherent risk measures Andy Philpott Vitor de Matos y Erlon Finardi z August 13, 2012 Abstract We consider a class of multistage stochastic linear programs
More informationMultistage risk-averse asset allocation with transaction costs
Multistage risk-averse asset allocation with transaction costs 1 Introduction Václav Kozmík 1 Abstract. This paper deals with asset allocation problems formulated as multistage stochastic programming models.
More informationRobust Dual Dynamic Programming
1 / 18 Robust Dual Dynamic Programming Angelos Georghiou, Angelos Tsoukalas, Wolfram Wiesemann American University of Beirut Olayan School of Business 31 May 217 2 / 18 Inspired by SDDP Stochastic optimization
More informationMartingale Pricing Theory in Discrete-Time and Discrete-Space Models
IEOR E4707: Foundations of Financial Engineering c 206 by Martin Haugh Martingale Pricing Theory in Discrete-Time and Discrete-Space Models These notes develop the theory of martingale pricing in a discrete-time,
More informationReport for technical cooperation between Georgia Institute of Technology and ONS - Operador Nacional do Sistema Elétrico Risk Averse Approach
Report for technical cooperation between Georgia Institute of Technology and ONS - Operador Nacional do Sistema Elétrico Risk Averse Approach Alexander Shapiro and Wajdi Tekaya School of Industrial and
More informationCourse notes for EE394V Restructured Electricity Markets: Locational Marginal Pricing
Course notes for EE394V Restructured Electricity Markets: Locational Marginal Pricing Ross Baldick Copyright c 2018 Ross Baldick www.ece.utexas.edu/ baldick/classes/394v/ee394v.html Title Page 1 of 160
More informationMultistage Stochastic Demand-side Management for Price-Making Major Consumers of Electricity in a Co-optimized Energy and Reserve Market
Multistage Stochastic Demand-side Management for Price-Making Major Consumers of Electricity in a Co-optimized Energy and Reserve Market Mahbubeh Habibian Anthony Downward Golbon Zakeri Abstract In this
More informationModels for estimating the performance of electricity markets with hydroelectric reservoir storage
Models for estimating the performance of electricity markets with hydroelectric reservoir storage Andy Philpott Ziming Guan May 12, 2013 Abstract We describe some new results of an empirical study of the
More informationDynamic sampling algorithms for multi-stage stochastic programs with risk aversion
Dynamic sampling algorithms for multi-stage stochastic programs with risk aversion A.B. Philpott y and V.L. de Matos z October 7, 2011 Abstract We consider the incorporation of a time-consistent coherent
More informationFinancial Transmission Rights Markets: An Overview
Financial Transmission Rights Markets: An Overview Golbon Zakeri A. Downward Department of Engineering Science, University of Auckland October 26, 2010 Outline Introduce financial transmission rights (FTRs).
More informationEquilibrium, uncertainty and risk in hydro-thermal electricity systems
Equilibrium, uncertainty and risk in hydro-thermal electricity systems Andy Philpott Michael Ferris Roger Wets August 31, 2015 Abstract The correspondence of competitive partial equilibrium with a social
More informationScenario reduction and scenario tree construction for power management problems
Scenario reduction and scenario tree construction for power management problems N. Gröwe-Kuska, H. Heitsch and W. Römisch Humboldt-University Berlin Institute of Mathematics Page 1 of 20 IEEE Bologna POWER
More informationElectricity markets, perfect competition and energy shortage risks
lectric ower ptimization entre lectricity markets, perfect competition and energy shortage risks Andy hilpott lectric ower ptimization entre University of Auckland http://www.epoc.org.nz joint work with
More informationDynamic sampling algorithms for multi-stage stochastic programs with risk aversion
Dynamic sampling algorithms for multi-stage stochastic programs with risk aversion A.B. Philpott y and V.L. de Matos z March 28, 2011 Abstract We consider the incorporation of a time-consistent coherent
More informationChapter 7: Portfolio Theory
Chapter 7: Portfolio Theory 1. Introduction 2. Portfolio Basics 3. The Feasible Set 4. Portfolio Selection Rules 5. The Efficient Frontier 6. Indifference Curves 7. The Two-Asset Portfolio 8. Unrestriceted
More informationOnline Appendix: Extensions
B Online Appendix: Extensions In this online appendix we demonstrate that many important variations of the exact cost-basis LUL framework remain tractable. In particular, dual problem instances corresponding
More informationStochastic Dual Dynamic Programming
1 / 43 Stochastic Dual Dynamic Programming Operations Research Anthony Papavasiliou 2 / 43 Contents [ 10.4 of BL], [Pereira, 1991] 1 Recalling the Nested L-Shaped Decomposition 2 Drawbacks of Nested Decomposition
More informationSOLVING ROBUST SUPPLY CHAIN PROBLEMS
SOLVING ROBUST SUPPLY CHAIN PROBLEMS Daniel Bienstock Nuri Sercan Özbay Columbia University, New York November 13, 2005 Project with Lucent Technologies Optimize the inventory buffer levels in a complicated
More informationDynamic Portfolio Choice II
Dynamic Portfolio Choice II Dynamic Programming Leonid Kogan MIT, Sloan 15.450, Fall 2010 c Leonid Kogan ( MIT, Sloan ) Dynamic Portfolio Choice II 15.450, Fall 2010 1 / 35 Outline 1 Introduction to Dynamic
More information1 Consumption and saving under uncertainty
1 Consumption and saving under uncertainty 1.1 Modelling uncertainty As in the deterministic case, we keep assuming that agents live for two periods. The novelty here is that their earnings in the second
More informationEE266 Homework 5 Solutions
EE, Spring 15-1 Professor S. Lall EE Homework 5 Solutions 1. A refined inventory model. In this problem we consider an inventory model that is more refined than the one you ve seen in the lectures. The
More informationDASC: A DECOMPOSITION ALGORITHM FOR MULTISTAGE STOCHASTIC PROGRAMS WITH STRONGLY CONVEX COST FUNCTIONS
DASC: A DECOMPOSITION ALGORITHM FOR MULTISTAGE STOCHASTIC PROGRAMS WITH STRONGLY CONVEX COST FUNCTIONS Vincent Guigues School of Applied Mathematics, FGV Praia de Botafogo, Rio de Janeiro, Brazil vguigues@fgv.br
More information1 Precautionary Savings: Prudence and Borrowing Constraints
1 Precautionary Savings: Prudence and Borrowing Constraints In this section we study conditions under which savings react to changes in income uncertainty. Recall that in the PIH, when you abstract from
More informationCharacterization of the Optimum
ECO 317 Economics of Uncertainty Fall Term 2009 Notes for lectures 5. Portfolio Allocation with One Riskless, One Risky Asset Characterization of the Optimum Consider a risk-averse, expected-utility-maximizing
More informationMonetary Economics Final Exam
316-466 Monetary Economics Final Exam 1. Flexible-price monetary economics (90 marks). Consider a stochastic flexibleprice money in the utility function model. Time is discrete and denoted t =0, 1,...
More informationCourse notes for EE394V Restructured Electricity Markets: Locational Marginal Pricing
Course notes for EE394V Restructured Electricity Markets: Locational Marginal Pricing Ross Baldick Copyright c 2017 Ross Baldick www.ece.utexas.edu/ baldick/classes/394v/ee394v.html Title Page 1 of 205
More informationStochastic Dual Dynamic Programming Algorithm for Multistage Stochastic Programming
Stochastic Dual Dynamic Programg Algorithm for Multistage Stochastic Programg Final presentation ISyE 8813 Fall 2011 Guido Lagos Wajdi Tekaya Georgia Institute of Technology November 30, 2011 Multistage
More informationInteger Programming Models
Integer Programming Models Fabio Furini December 10, 2014 Integer Programming Models 1 Outline 1 Combinatorial Auctions 2 The Lockbox Problem 3 Constructing an Index Fund Integer Programming Models 2 Integer
More informationEssays on Some Combinatorial Optimization Problems with Interval Data
Essays on Some Combinatorial Optimization Problems with Interval Data a thesis submitted to the department of industrial engineering and the institute of engineering and sciences of bilkent university
More informationLECTURE 2: MULTIPERIOD MODELS AND TREES
LECTURE 2: MULTIPERIOD MODELS AND TREES 1. Introduction One-period models, which were the subject of Lecture 1, are of limited usefulness in the pricing and hedging of derivative securities. In real-world
More informationEC316a: Advanced Scientific Computation, Fall Discrete time, continuous state dynamic models: solution methods
EC316a: Advanced Scientific Computation, Fall 2003 Notes Section 4 Discrete time, continuous state dynamic models: solution methods We consider now solution methods for discrete time models in which decisions
More informationEnergy Systems under Uncertainty: Modeling and Computations
Energy Systems under Uncertainty: Modeling and Computations W. Römisch Humboldt-University Berlin Department of Mathematics www.math.hu-berlin.de/~romisch Systems Analysis 2015, November 11 13, IIASA (Laxenburg,
More informationHandout 8: Introduction to Stochastic Dynamic Programming. 2 Examples of Stochastic Dynamic Programming Problems
SEEM 3470: Dynamic Optimization and Applications 2013 14 Second Term Handout 8: Introduction to Stochastic Dynamic Programming Instructor: Shiqian Ma March 10, 2014 Suggested Reading: Chapter 1 of Bertsekas,
More information16 MAKING SIMPLE DECISIONS
247 16 MAKING SIMPLE DECISIONS Let us associate each state S with a numeric utility U(S), which expresses the desirability of the state A nondeterministic action A will have possible outcome states Result
More information4: SINGLE-PERIOD MARKET MODELS
4: SINGLE-PERIOD MARKET MODELS Marek Rutkowski School of Mathematics and Statistics University of Sydney Semester 2, 2016 M. Rutkowski (USydney) Slides 4: Single-Period Market Models 1 / 87 General Single-Period
More informationFinancial Optimization ISE 347/447. Lecture 15. Dr. Ted Ralphs
Financial Optimization ISE 347/447 Lecture 15 Dr. Ted Ralphs ISE 347/447 Lecture 15 1 Reading for This Lecture C&T Chapter 12 ISE 347/447 Lecture 15 2 Stock Market Indices A stock market index is a statistic
More informationWhat can we do with numerical optimization?
Optimization motivation and background Eddie Wadbro Introduction to PDE Constrained Optimization, 2016 February 15 16, 2016 Eddie Wadbro, Introduction to PDE Constrained Optimization, February 15 16, 2016
More informationEco504 Spring 2010 C. Sims FINAL EXAM. β t 1 2 φτ2 t subject to (1)
Eco54 Spring 21 C. Sims FINAL EXAM There are three questions that will be equally weighted in grading. Since you may find some questions take longer to answer than others, and partial credit will be given
More informationEcon 101A Final Exam We May 9, 2012.
Econ 101A Final Exam We May 9, 2012. You have 3 hours to answer the questions in the final exam. We will collect the exams at 2.30 sharp. Show your work, and good luck! Problem 1. Utility Maximization.
More information1 Dynamic programming
1 Dynamic programming A country has just discovered a natural resource which yields an income per period R measured in terms of traded goods. The cost of exploitation is negligible. The government wants
More informationOptimal Dam Management
Optimal Dam Management Michel De Lara et Vincent Leclère July 3, 2012 Contents 1 Problem statement 1 1.1 Dam dynamics.................................. 2 1.2 Intertemporal payoff criterion..........................
More informationStochastic Optimization
Stochastic Optimization Introduction and Examples Alireza Ghaffari-Hadigheh Azarbaijan Shahid Madani University (ASMU) hadigheha@azaruniv.edu Fall 2017 Alireza Ghaffari-Hadigheh (ASMU) Stochastic Optimization
More informationA Robust Option Pricing Problem
IMA 2003 Workshop, March 12-19, 2003 A Robust Option Pricing Problem Laurent El Ghaoui Department of EECS, UC Berkeley 3 Robust optimization standard form: min x sup u U f 0 (x, u) : u U, f i (x, u) 0,
More information3.2 No-arbitrage theory and risk neutral probability measure
Mathematical Models in Economics and Finance Topic 3 Fundamental theorem of asset pricing 3.1 Law of one price and Arrow securities 3.2 No-arbitrage theory and risk neutral probability measure 3.3 Valuation
More informationDRAFT. 1 exercise in state (S, t), π(s, t) = 0 do not exercise in state (S, t) Review of the Risk Neutral Stock Dynamics
Chapter 12 American Put Option Recall that the American option has strike K and maturity T and gives the holder the right to exercise at any time in [0, T ]. The American option is not straightforward
More informationLecture 17: More on Markov Decision Processes. Reinforcement learning
Lecture 17: More on Markov Decision Processes. Reinforcement learning Learning a model: maximum likelihood Learning a value function directly Monte Carlo Temporal-difference (TD) learning COMP-424, Lecture
More informationCourse Handouts - Introduction ECON 8704 FINANCIAL ECONOMICS. Jan Werner. University of Minnesota
Course Handouts - Introduction ECON 8704 FINANCIAL ECONOMICS Jan Werner University of Minnesota SPRING 2019 1 I.1 Equilibrium Prices in Security Markets Assume throughout this section that utility functions
More informationOptimal Security Liquidation Algorithms
Optimal Security Liquidation Algorithms Sergiy Butenko Department of Industrial Engineering, Texas A&M University, College Station, TX 77843-3131, USA Alexander Golodnikov Glushkov Institute of Cybernetics,
More informationPh.D. Preliminary Examination MICROECONOMIC THEORY Applied Economics Graduate Program August 2017
Ph.D. Preliminary Examination MICROECONOMIC THEORY Applied Economics Graduate Program August 2017 The time limit for this exam is four hours. The exam has four sections. Each section includes two questions.
More informationSIMULATION OF ELECTRICITY MARKETS
SIMULATION OF ELECTRICITY MARKETS MONTE CARLO METHODS Lectures 15-18 in EG2050 System Planning Mikael Amelin 1 COURSE OBJECTIVES To pass the course, the students should show that they are able to - apply
More informationDynamic Replication of Non-Maturing Assets and Liabilities
Dynamic Replication of Non-Maturing Assets and Liabilities Michael Schürle Institute for Operations Research and Computational Finance, University of St. Gallen, Bodanstr. 6, CH-9000 St. Gallen, Switzerland
More informationArbitrage Conditions for Electricity Markets with Production and Storage
SWM ORCOS Arbitrage Conditions for Electricity Markets with Production and Storage Raimund Kovacevic Research Report 2018-03 March 2018 ISSN 2521-313X Operations Research and Control Systems Institute
More informationRisk aversion in multi-stage stochastic programming: a modeling and algorithmic perspective
Risk aversion in multi-stage stochastic programming: a modeling and algorithmic perspective Tito Homem-de-Mello School of Business Universidad Adolfo Ibañez, Santiago, Chile Joint work with Bernardo Pagnoncelli
More informationBenchmarking wholesale hydroelectricity markets with risk-averse agents
Benchmarking wholesale hydroelectricity markets with risk-averse agents Andy Philpott Ziming Guan November 1, 2018 Abstract Using a suite of empirical models we study the effects of randomness and risk
More informationLecture 7: Bayesian approach to MAB - Gittins index
Advanced Topics in Machine Learning and Algorithmic Game Theory Lecture 7: Bayesian approach to MAB - Gittins index Lecturer: Yishay Mansour Scribe: Mariano Schain 7.1 Introduction In the Bayesian approach
More informationMicroeconomics of Banking: Lecture 2
Microeconomics of Banking: Lecture 2 Prof. Ronaldo CARPIO September 25, 2015 A Brief Look at General Equilibrium Asset Pricing Last week, we saw a general equilibrium model in which banks were irrelevant.
More informationDynamic Portfolio Execution Detailed Proofs
Dynamic Portfolio Execution Detailed Proofs Gerry Tsoukalas, Jiang Wang, Kay Giesecke March 16, 2014 1 Proofs Lemma 1 (Temporary Price Impact) A buy order of size x being executed against i s ask-side
More informationMarkov Decision Processes: Making Decision in the Presence of Uncertainty. (some of) R&N R&N
Markov Decision Processes: Making Decision in the Presence of Uncertainty (some of) R&N 16.1-16.6 R&N 17.1-17.4 Different Aspects of Machine Learning Supervised learning Classification - concept learning
More informationPh.D. Preliminary Examination MICROECONOMIC THEORY Applied Economics Graduate Program June 2017
Ph.D. Preliminary Examination MICROECONOMIC THEORY Applied Economics Graduate Program June 2017 The time limit for this exam is four hours. The exam has four sections. Each section includes two questions.
More informationUncertainty in Equilibrium
Uncertainty in Equilibrium Larry Blume May 1, 2007 1 Introduction The state-preference approach to uncertainty of Kenneth J. Arrow (1953) and Gérard Debreu (1959) lends itself rather easily to Walrasian
More informationChallenges and Solutions: Innovations that we need in Optimization for Future Electric Power Systems
Challenges and Solutions: Innovations that we need in Optimization for Future Electric Power Systems Dr. Chenye Wu Prof. Gabriela Hug wu@eeh.ee.ethz.ch 1 Challenges in the traditional power systems The
More informationMATH 5510 Mathematical Models of Financial Derivatives. Topic 1 Risk neutral pricing principles under single-period securities models
MATH 5510 Mathematical Models of Financial Derivatives Topic 1 Risk neutral pricing principles under single-period securities models 1.1 Law of one price and Arrow securities 1.2 No-arbitrage theory and
More informationContents Critique 26. portfolio optimization 32
Contents Preface vii 1 Financial problems and numerical methods 3 1.1 MATLAB environment 4 1.1.1 Why MATLAB? 5 1.2 Fixed-income securities: analysis and portfolio immunization 6 1.2.1 Basic valuation of
More informationLecture 8: Introduction to asset pricing
THE UNIVERSITY OF SOUTHAMPTON Paul Klein Office: Murray Building, 3005 Email: p.klein@soton.ac.uk URL: http://paulklein.se Economics 3010 Topics in Macroeconomics 3 Autumn 2010 Lecture 8: Introduction
More informationCrediting Wind and Solar Renewables in Electricity Capacity Markets: The Effects of Alternative Definitions upon Market Efficiency. The Energy Journal
Crediting Wind and Solar Renewables in Electricity Capacity Markets: The Effects of Alternative Definitions upon Market Efficiency The Energy Journal On-Line Appendix A: Supporting proofs of social cost
More informationTDT4171 Artificial Intelligence Methods
TDT47 Artificial Intelligence Methods Lecture 7 Making Complex Decisions Norwegian University of Science and Technology Helge Langseth IT-VEST 0 helgel@idi.ntnu.no TDT47 Artificial Intelligence Methods
More informationResource Planning with Uncertainty for NorthWestern Energy
Resource Planning with Uncertainty for NorthWestern Energy Selection of Optimal Resource Plan for 213 Resource Procurement Plan August 28, 213 Gary Dorris, Ph.D. Ascend Analytics, LLC gdorris@ascendanalytics.com
More informationMicroeconomics II. CIDE, MsC Economics. List of Problems
Microeconomics II CIDE, MsC Economics List of Problems 1. There are three people, Amy (A), Bart (B) and Chris (C): A and B have hats. These three people are arranged in a room so that B can see everything
More informationEquity correlations implied by index options: estimation and model uncertainty analysis
1/18 : estimation and model analysis, EDHEC Business School (joint work with Rama COT) Modeling and managing financial risks Paris, 10 13 January 2011 2/18 Outline 1 2 of multi-asset models Solution to
More informationB. Online Appendix. where ɛ may be arbitrarily chosen to satisfy 0 < ɛ < s 1 and s 1 is defined in (B1). This can be rewritten as
B Online Appendix B1 Constructing examples with nonmonotonic adoption policies Assume c > 0 and the utility function u(w) is increasing and approaches as w approaches 0 Suppose we have a prior distribution
More informationMAFS Computational Methods for Pricing Structured Products
MAFS550 - Computational Methods for Pricing Structured Products Solution to Homework Two Course instructor: Prof YK Kwok 1 Expand f(x 0 ) and f(x 0 x) at x 0 into Taylor series, where f(x 0 ) = f(x 0 )
More informationSubmission on. Consultation Paper. Managing locational price risk proposal. Electric Power Optimization Centre. University of Auckland
Submission on P Consultation Paper E O Managing locational price risk proposal C by Electric Power Optimization Centre University of Auckland Professor Andy Philpott Dr Golbon Zakeri Dr Geoff Pritchard
More informationPart 1: q Theory and Irreversible Investment
Part 1: q Theory and Irreversible Investment Goal: Endogenize firm characteristics and risk. Value/growth Size Leverage New issues,... This lecture: q theory of investment Irreversible investment and real
More informationChapter 9 Dynamic Models of Investment
George Alogoskoufis, Dynamic Macroeconomic Theory, 2015 Chapter 9 Dynamic Models of Investment In this chapter we present the main neoclassical model of investment, under convex adjustment costs. This
More informationProblem Set 3. Thomas Philippon. April 19, Human Wealth, Financial Wealth and Consumption
Problem Set 3 Thomas Philippon April 19, 2002 1 Human Wealth, Financial Wealth and Consumption The goal of the question is to derive the formulas on p13 of Topic 2. This is a partial equilibrium analysis
More informationHedging with Life and General Insurance Products
Hedging with Life and General Insurance Products June 2016 2 Hedging with Life and General Insurance Products Jungmin Choi Department of Mathematics East Carolina University Abstract In this study, a hybrid
More informationProgressive Hedging for Multi-stage Stochastic Optimization Problems
Progressive Hedging for Multi-stage Stochastic Optimization Problems David L. Woodruff Jean-Paul Watson Graduate School of Management University of California, Davis Davis, CA 95616, USA dlwoodruff@ucdavis.edu
More information1 Shapley-Shubik Model
1 Shapley-Shubik Model There is a set of buyers B and a set of sellers S each selling one unit of a good (could be divisible or not). Let v ij 0 be the monetary value that buyer j B assigns to seller i
More informationSolving dynamic portfolio choice problems by recursing on optimized portfolio weights or on the value function?
DOI 0.007/s064-006-9073-z ORIGINAL PAPER Solving dynamic portfolio choice problems by recursing on optimized portfolio weights or on the value function? Jules H. van Binsbergen Michael W. Brandt Received:
More informationPayment mechanisms and risk-aversion in electricity markets with uncertain supply
Payment mechanisms and risk-aversion in electricity markets with uncertain supply Ryan Cory-Wright Joint work with Golbon Zakeri (thanks to Andy Philpott) ISMP, Bordeaux, July 2018. ORC, Massachusetts
More informationROBUST OPTIMIZATION OF MULTI-PERIOD PRODUCTION PLANNING UNDER DEMAND UNCERTAINTY. A. Ben-Tal, B. Golany and M. Rozenblit
ROBUST OPTIMIZATION OF MULTI-PERIOD PRODUCTION PLANNING UNDER DEMAND UNCERTAINTY A. Ben-Tal, B. Golany and M. Rozenblit Faculty of Industrial Engineering and Management, Technion, Haifa 32000, Israel ABSTRACT
More informationDepartment of Agricultural Economics. PhD Qualifier Examination. August 2010
Department of Agricultural Economics PhD Qualifier Examination August 200 Instructions: The exam consists of six questions. You must answer all questions. If you need an assumption to complete a question,
More informationExtraction capacity and the optimal order of extraction. By: Stephen P. Holland
Extraction capacity and the optimal order of extraction By: Stephen P. Holland Holland, Stephen P. (2003) Extraction Capacity and the Optimal Order of Extraction, Journal of Environmental Economics and
More informationMultistage Stochastic Programming
IE 495 Lecture 21 Multistage Stochastic Programming Prof. Jeff Linderoth April 16, 2003 April 16, 2002 Stochastic Programming Lecture 21 Slide 1 Outline HW Fixes Multistage Stochastic Programming Modeling
More informationThe application of linear programming to management accounting
The application of linear programming to management accounting After studying this chapter, you should be able to: formulate the linear programming model and calculate marginal rates of substitution and
More information16 MAKING SIMPLE DECISIONS
253 16 MAKING SIMPLE DECISIONS Let us associate each state S with a numeric utility U(S), which expresses the desirability of the state A nondeterministic action a will have possible outcome states Result(a)
More informationRevenue Management Under the Markov Chain Choice Model
Revenue Management Under the Markov Chain Choice Model Jacob B. Feldman School of Operations Research and Information Engineering, Cornell University, Ithaca, New York 14853, USA jbf232@cornell.edu Huseyin
More informationThe Uncertain Volatility Model
The Uncertain Volatility Model Claude Martini, Antoine Jacquier July 14, 008 1 Black-Scholes and realised volatility What happens when a trader uses the Black-Scholes (BS in the sequel) formula to sell
More informationIntroduction to modeling using stochastic programming. Andy Philpott The University of Auckland
Introduction to modeling using stochastic programming Andy Philpott The University of Auckland Tutorial presentation at SPX, Tuscon, October 9th, 2004 Summary Introduction to basic concepts Risk Multi-stage
More informationStochastic Programming and Financial Analysis IE447. Midterm Review. Dr. Ted Ralphs
Stochastic Programming and Financial Analysis IE447 Midterm Review Dr. Ted Ralphs IE447 Midterm Review 1 Forming a Mathematical Programming Model The general form of a mathematical programming model is:
More informationOnline Appendix Optimal Time-Consistent Government Debt Maturity D. Debortoli, R. Nunes, P. Yared. A. Proofs
Online Appendi Optimal Time-Consistent Government Debt Maturity D. Debortoli, R. Nunes, P. Yared A. Proofs Proof of Proposition 1 The necessity of these conditions is proved in the tet. To prove sufficiency,
More informationOn the Optimality of Financial Repression
On the Optimality of Financial Repression V.V. Chari, Alessandro Dovis and Patrick Kehoe Conference in honor of Robert E. Lucas Jr, October 2016 Financial Repression Regulation forcing financial institutions
More informationThe Sharing Economy for the Smart Grid
The Sharing Economy for the Smart Grid Kameshwar Poolla UC Berkeley PSERC Webinar February 7, 2017 February 7, 2017 0 / 31 Shared Electricity Services The New Sharing Economy cars, homes, services,...
More information6.231 DYNAMIC PROGRAMMING LECTURE 10 LECTURE OUTLINE
6.231 DYNAMIC PROGRAMMING LECTURE 10 LECTURE OUTLINE Rollout algorithms Cost improvement property Discrete deterministic problems Approximations of rollout algorithms Discretization of continuous time
More information17 MAKING COMPLEX DECISIONS
267 17 MAKING COMPLEX DECISIONS The agent s utility now depends on a sequence of decisions In the following 4 3grid environment the agent makes a decision to move (U, R, D, L) at each time step When the
More informationLecture 2: Making Good Sequences of Decisions Given a Model of World. CS234: RL Emma Brunskill Winter 2018
Lecture 2: Making Good Sequences of Decisions Given a Model of World CS234: RL Emma Brunskill Winter 218 Human in the loop exoskeleton work from Steve Collins lab Class Structure Last Time: Introduction
More informationForward Contracts and Generator Market Power: How Externalities Reduce Benefits in Equilibrium
Forward Contracts and Generator Market Power: How Externalities Reduce Benefits in Equilibrium Ian Schneider, Audun Botterud, and Mardavij Roozbehani November 9, 2017 Abstract Research has shown that forward
More informationDynamic Risk Management in Electricity Portfolio Optimization via Polyhedral Risk Functionals
Dynamic Risk Management in Electricity Portfolio Optimization via Polyhedral Risk Functionals A. Eichhorn and W. Römisch Humboldt-University Berlin, Department of Mathematics, Germany http://www.math.hu-berlin.de/~romisch
More information